Cursor’s New Router Puts a Price Tag on Every AI Coding Decision

Cursor, the AI coding assistant owned by SpaceX after a $60 billion acquisition, launched Cursor Router, a feature that automatically assigns each coding request to the most cost-effective AI model, reporting cost reductions of 30% to 50% in early enterprise use and up to 60% in broader testing. The router, trained on over 600,000 live requests and tested on millions more, offers three modes—Intelligence, Balance, and Cost—to help teams manage AI coding expenses, with per-commit costs ranging from $4.63 to $6.76 compared to $7.34 for Opus 4.8 and $12.69 for Fable 5.

TL;DR — Key Takeaways Cursor is automating model selection: Its router analyzes each coding request and assigns it to the model best suited to the task. The potential savings are substantial: Cursor reports cost reductions of 30% to 50% in early enterprise use and up to 60% in broader testing. Teams can choose their priority: Intelligence, Balance and Cost modes let administrators trade off model capability against spending. AI coding costs now need governance: Automatically using premium models for routine work can create significant unnecessary expense at scale. Cursor, the AI coding assistant now owned by SpaceX after last month’s $60 billion acquisition, rolled out a new feature this week that answers a question a lot of engineering leaders have been quietly avoiding: which AI model should actually handle this task? The answer, until now, has mostly been “whichever one the developer likes.” Cursor says roughly 60% of its users pick a single model and stick with it for everything, from fixing a typo to untangling a gnarly refactor. That habit is comfortable, but it’s expensive. Frontier models cost frontier prices, whether the job in front of them needs that horsepower or not. Cursor Router is built to break that habit. It’s a classifier that sits in front of every request and decides, automatically, which model should handle it. Simple, routine work gets sent to cheaper models. Complex, long-horizon problems go to the frontier reasoning models built for them. The system considers the query itself, the surrounding context, task complexity, and what Cursor has learned about each model’s strengths. That last part is where Cursor has an edge most vendors don’t. The company says it processes hundreds of millions of coding requests every week, giving it a direct view into what actually works and what code developers keep versus what they throw out. Cursor trained the router on more than 600,000 live requests and then tested it in production across millions more, rather than relying purely on offline benchmarks. That’s a meaningful distinction. Offline evals are useful, but they miss real-world friction points, like the extra cost that comes from switching models mid-conversation. Cursor’s evaluation accounts for that. The results, at least according to Cursor’s own numbers, are hard to ignore. During early access with a handful of enterprise customers, teams saw costs drop 30% to 50% compared to running everything through Opus 4.8, with no drop in output quality. In broader A/B testing across millions of requests, the company reported savings as high as 60%. Looking at cost per commit rather than cost per request, Cursor Router landed at $6.76 for its Intelligence mode and $4.63 for Balance mode, compared to $7.34 for Opus 4.8 and $12.69 for Fable 5 running solo. Cursor Router offers three modes: Intelligence, Balance and Cost. Each one moves a team along what Cursor calls the cost-intelligence Pareto frontier. Intelligence mode aims for frontier-level output regardless of price. Cost mode optimizes token spend while still reaching for the highest intelligence it can afford. Balance sits in the middle, matching the performance most developers already reach for as a daily driver. Admins can control which modes their teams can access, set defaults, and allow or block specific underlying models. The feature is live now across desktop, web, iOS, the CLI, and Cursor’s SDK, and it’s on by default for Teams plans. The bigger story here isn’t really about Cursor. It’s about where AI coding tools are heading generally. Model selection has quietly become a budget line item, and most organizations don’t have a disciplined way to manage it. A developer defaulting to the priciest model available, out of habit rather than need, is the AI equivalent of leaving the lights on in an empty office. It doesn’t feel like a waste in the moment, but it adds up fast across thousands of developers making that choice thousands of times a day. Routing isn’t a new concept. Tools like OpenRouter have offered multi-provider routing at the API layer for a while, typically asking developers to configure their own fallback chains and provider preferences. Cursor’s approach is narrower and more opinionated by design. It’s built directly into the IDE, trained on Cursor’s own usage data, and aimed at teams who’d rather not configure anything at all. Whether that trade-off- less control in exchange for less friction- appeals to enterprise buyers will likely shape how much traction this feature gets outside of Cursor’s existing customer base. For SpaceX, the launch also marks the first real product move since the Cursor acquisition closed, giving the company something concrete to point to as it works out how Cursor fits into its broader AI ambitions. For everyone else building software with AI assistance, it’s another sign that the conversation is shifting from “which model is best” to “which model is worth it, for this task, right now.” That’s a more mature question, and one enterprise IT teams are going to need better tools to answer.